Learning Invariant Representations of Molecules for Atomization Energy Prediction